Skip to content

Latest commit

 

History

History
38 lines (30 loc) · 990 Bytes

README.md

File metadata and controls

38 lines (30 loc) · 990 Bytes
Raw

NeuralNet Document Classifier

Trains a model capable of classifying any document using some Huggingface transformers and LayoutXLM multi-language weights.

Requires:

  • PyTorch
  • Huggingface bindings
  • Tesseract installation as well as PyTesseract bindings for testing purposes

plus some pip dependencies.

To train the model, create a /train_data subdirectory on main.py dir level containing folders representing the classifier labels, e.G:

/invoices
 - invoice_train1.jpg
 - invoice_train2.png
 - invoice_train3.tif
/vehicle_registrations
 - vehicle_registration_train1.jpg
 - ...

and so forth

and run main.py. This will put out a model capable of being loaded and inferenced from by loader.py, which spins up a simple webserver serving the classified response in a form of:

{
    "class": "vehicle_registration",
    "class_id": 15,
    "confidence": "99.7821569442749",
    "inference_time": 0.055999755859375,
    "tokenizer_time": 1.834416389465332
}

~